RetroSearch Browse

Home - News ( United States | United Kingdom | Italy | Germany ) - Football scores

Showing content from https://learn.microsoft.com/en-us/dotnet/core/diagnostics/event-counters below:

EventCounters in .NET Core - .NET

This article applies to: âï¸ .NET Core 3.0 SDK and later versions

Note

For developing new .NET projects Microsoft recommends using the newer System.Diagnostics.Metrics APIs instead. The System.Diagnostics.Metrics APIs offer increased functionality, standardization and integration with a broader ecosystem of tools. See the metrics API comparison for more information.

EventCounters are .NET APIs used for lightweight, cross-platform, and near real-time performance metric collection. EventCounters were added as a cross-platform alternative to the "performance counters" of .NET Framework on Windows. In this article, you'll learn what EventCounters are, how to implement them, and how to consume them.

The .NET runtime and a few .NET libraries publish basic diagnostics information using EventCounters starting in .NET Core 3.0. Apart from the EventCounters that are provided by the .NET runtime, you may choose to implement your own EventCounters. EventCounters can be used to track various metrics. Learn more about them in well-known EventCounters in .NET.

EventCounters live as a part of an EventSource, and are automatically pushed to listener tools on a regular basis. Like all other events on an EventSource, they can be consumed both in-proc and out-of-proc via EventListener and EventPipe. This article focuses on the cross-platform capabilities of EventCounters, and intentionally excludes PerfView and ETW (Event Tracing for Windows) - although both can be used with EventCounters.

EventCounter API overview

There are two primary categories of EventCounters. Some counters are for "rate" values, such as total number of exceptions, total number of GCs, and total number of requests. Other counters are "snapshot" values, such as heap usage, CPU usage, and working set size. Within each of these categories of counters, there are two types of counters that vary by how they get their value. Polling counters retrieve their value via a callback, and non-polling counters have their values directly set on the counter instance.

The counters are represented by the following implementations:

An event listener specifies how long measurement intervals are. At the end of each interval a value is transmitted to the listener for each counter. The implementations of a counter determine what APIs and calculations are used to produce the value each interval.

The EventCounter records a set of values. The EventCounter.WriteMetric method adds a new value to the set. With each interval, a statistical summary for the set is computed, such as the min, max, and mean. The dotnet-counters tool will always display the mean value. The EventCounter is useful to describe a discrete set of operations. Common usage may include monitoring the average size in bytes of recent IO operations, or the average monetary value of a set of financial transactions.
The IncrementingEventCounter records a running total for each time interval. The IncrementingEventCounter.Increment method adds to the total. For example, if Increment() is called three times during one interval with values 1, 2, and 5, then the running total of 8 will be reported as the counter value for this interval. The dotnet-counters tool will display the rate as the recorded total / time. The IncrementingEventCounter is useful to measure how frequently an action is occurring, such as the number of requests processed per second.
The PollingCounter uses a callback to determine the value that is reported. With each time interval, the user provided callback function is invoked and the return value is used as the counter value. A PollingCounter can be used to query a metric from an external source, for example getting the current free bytes on a disk. It can also be used to report custom statistics that can be computed on demand by an application. Examples include reporting the 95th percentile of recent request latencies, or the current hit or miss ratio of a cache.
The IncrementingPollingCounter uses a callback to determine the reported increment value. With each time interval, the callback is invoked, and then the difference between the current invocation, and the last invocation is the reported value. The dotnet-counters tool will always display the difference as a rate, the reported value / time. This counter is useful when it isn't feasible to call an API on each occurrence, but it's possible to query the total number of occurrences. For example, you could report the number of bytes written to a file per second, even without a notification each time a byte is written.

Implement an EventSource

The following code implements a sample EventSource exposed as the named "Sample.EventCounter.Minimal" provider. This source contains an EventCounter representing request-processing time. Such a counter has a name (that is, its unique ID in the source) and a display name, both used by listener tools such as dotnet-counters.

using System.Diagnostics.Tracing;

[EventSource(Name = "Sample.EventCounter.Minimal")]
public sealed class MinimalEventCounterSource : EventSource
{
    public static readonly MinimalEventCounterSource Log = new MinimalEventCounterSource();

    private EventCounter _requestCounter;

    private MinimalEventCounterSource() =>
        _requestCounter = new EventCounter("request-time", this)
        {
            DisplayName = "Request Processing Time",
            DisplayUnits = "ms"
        };

    public void Request(string url, long elapsedMilliseconds)
    {
        WriteEvent(1, url, elapsedMilliseconds);
        _requestCounter?.WriteMetric(elapsedMilliseconds);
    }

    protected override void Dispose(bool disposing)
    {
        _requestCounter?.Dispose();
        _requestCounter = null;

        base.Dispose(disposing);
    }
}

You use dotnet-counters ps to display a list of .NET processes that can be monitored:

dotnet-counters ps
   1398652 dotnet     C:\Program Files\dotnet\dotnet.exe
   1399072 dotnet     C:\Program Files\dotnet\dotnet.exe
   1399112 dotnet     C:\Program Files\dotnet\dotnet.exe
   1401880 dotnet     C:\Program Files\dotnet\dotnet.exe
   1400180 sample-counters C:\sample-counters\bin\Debug\netcoreapp3.1\sample-counters.exe

Pass the EventSource name to the --counters option to start monitoring your counter:

dotnet-counters monitor --process-id 1400180 --counters Sample.EventCounter.Minimal

The following example shows monitor output:

Press p to pause, r to resume, q to quit.
    Status: Running

[Samples-EventCounterDemos-Minimal]
    Request Processing Time (ms)                            0.445

Press q to stop the monitoring command.

Conditional counters

When implementing an EventSource, the containing counters can be conditionally instantiated when the EventSource.OnEventCommand method is called with a Command value of EventCommand.Enable. To safely instantiate a counter instance only if it is null, use the null-coalescing assignment operator. Additionally, custom methods can evaluate the IsEnabled method to determine whether or not the current event source is enabled.

using System.Diagnostics.Tracing;

[EventSource(Name = "Sample.EventCounter.Conditional")]
public sealed class ConditionalEventCounterSource : EventSource
{
    public static readonly ConditionalEventCounterSource Log = new ConditionalEventCounterSource();

    private EventCounter _requestCounter;

    private ConditionalEventCounterSource() { }

    protected override void OnEventCommand(EventCommandEventArgs args)
    {
        if (args.Command == EventCommand.Enable)
        {
            _requestCounter ??= new EventCounter("request-time", this)
            {
                DisplayName = "Request Processing Time",
                DisplayUnits = "ms"
            };
        }
    }

    public void Request(string url, float elapsedMilliseconds)
    {
        if (IsEnabled())
        {
            _requestCounter?.WriteMetric(elapsedMilliseconds);
        }
    }

    protected override void Dispose(bool disposing)
    {
        _requestCounter?.Dispose();
        _requestCounter = null;

        base.Dispose(disposing);
    }
}

Tip

Conditional counters are counters that are conditionally instantiated, a micro-optimization. The runtime adopts this pattern for scenarios where counters are normally not used, to save a fraction of a millisecond.

.NET Core runtime example counters

There are many great example implementations in the .NET Core runtime. Here is the runtime implementation for the counter that tracks the working set size of the application.

var workingSetCounter = new PollingCounter(
    "working-set",
    this,
    () => (double)(Environment.WorkingSet / 1_000_000))
{
    DisplayName = "Working Set",
    DisplayUnits = "MB"
};

The PollingCounter reports the current amount of physical memory mapped to the process (working set) of the app, since it captures a metric at a moment in time. The callback for polling a value is the provided lambda expression, which is just a call to the System.Environment.WorkingSet API. DisplayName and DisplayUnits are optional properties that can be set to help the consumer side of the counter to display the value more clearly. For example, dotnet-counters uses these properties to display the more display-friendly version of the counter names.

Important

The DisplayName properties are not localized.

For the PollingCounter, and the IncrementingPollingCounter, nothing else needs to be done. They both poll the values themselves at an interval requested by the consumer.

Here is an example of a runtime counter implemented using IncrementingPollingCounter.

var monitorContentionCounter = new IncrementingPollingCounter(
    "monitor-lock-contention-count",
    this,
    () => Monitor.LockContentionCount
)
{
    DisplayName = "Monitor Lock Contention Count",
    DisplayRateTimeScale = TimeSpan.FromSeconds(1)
};

The IncrementingPollingCounter uses the Monitor.LockContentionCount API to report the increment of the total lock contention count. The DisplayRateTimeScale property is optional, but when used it can provide a hint for what time interval the counter is best displayed at. For example, the lock contention count is best displayed as count per second, so its DisplayRateTimeScale is set to one second. The display rate can be adjusted for different types of rate counters.

There are more counter implementations to use as a reference in the .NET runtime repo.

Concurrency

Tip

The EventCounters API does not guarantee thread safety. When the delegates passed to PollingCounter or IncrementingPollingCounter instances are called by multiple threads, it's your responsibility to guarantee the delegates' thread-safety.

For example, consider the following EventSource to keep track of requests.

using System;
using System.Diagnostics.Tracing;

public class RequestEventSource : EventSource
{
    public static readonly RequestEventSource Log = new RequestEventSource();

    private IncrementingPollingCounter _requestRateCounter;
    private long _requestCount = 0;

    private RequestEventSource() =>
        _requestRateCounter = new IncrementingPollingCounter("request-rate", this, () => _requestCount)
        {
            DisplayName = "Request Rate",
            DisplayRateTimeScale = TimeSpan.FromSeconds(1)
        };

    public void AddRequest() => ++ _requestCount;

    protected override void Dispose(bool disposing)
    {
        _requestRateCounter?.Dispose();
        _requestRateCounter = null;

        base.Dispose(disposing);
    }
}

The AddRequest() method can be called from a request handler, and the RequestRateCounter polls the value at the interval specified by the consumer of the counter. However, the AddRequest() method can be called by multiple threads at once, putting a race condition on _requestCount. A thread-safe alternative way to increment the _requestCount is to use Interlocked.Increment.

public void AddRequest() => Interlocked.Increment(ref _requestCount);

To prevent torn reads (on 32-bit architectures) of the long-field _requestCount use Interlocked.Read.

_requestRateCounter = new IncrementingPollingCounter("request-rate", this, () => Interlocked.Read(ref _requestCount))
{
    DisplayName = "Request Rate",
    DisplayRateTimeScale = TimeSpan.FromSeconds(1)
};

Consume EventCounters

There are two primary ways of consuming EventCounters: in-proc and out-of-proc. The consumption of EventCounters can be distinguished into three layers of various consuming technologies.

Transporting events in a raw stream via ETW or EventPipe:

ETW APIs come with the Windows OS, and EventPipe is accessible as a .NET API, or the diagnostic IPC protocol.
Decoding the binary event stream into events:

The TraceEvent library handles both ETW and EventPipe stream formats.
Command-line and GUI tools:

Tools like PerfView (ETW or EventPipe), dotnet-counters (EventPipe only), and dotnet-monitor (EventPipe only).

Consume out-of-proc

Consuming EventCounters out-of-proc is a common approach. You can use dotnet-counters to consume them in a cross-platform manner via an EventPipe. The dotnet-counters tool is a cross-platform dotnet CLI global tool that can be used to monitor the counter values. To find out how to use dotnet-counters to monitor your counters, see dotnet-counters, or work through the Measure performance using EventCounters tutorial.

Azure Application Insights

EventCounters can be consumed by Azure Monitor, specifically Azure Application Insights. Counters can be added and removed, and you're free to specify custom counters, or well-known counters. For more information, see Customizing counters to be collected.

dotnet-monitor

The dotnet-monitor tool makes it easier to access diagnostics from a .NET process in a remote and automated fashion. In addition to traces, it can monitor metrics, collect memory dumps, and collect GC dumps. It's distributed as both a CLI tool and a docker image. It exposes a REST API, and the collection of diagnostic artifacts occurs through REST calls.

For more information, see dotnet-monitor.

Consume in-proc

You can consume the counter values via the EventListener API. An EventListener is an in-proc way of consuming any events written by all instances of an EventSource in your application. For more information on how to use the EventListener API, see EventListener.

First, the EventSource that produces the counter value needs to be enabled. Override the EventListener.OnEventSourceCreated method to get a notification when an EventSource is created, and if this is the correct EventSource with your EventCounters, then you can call EventListener.EnableEvents on it. Here is an example override:

protected override void OnEventSourceCreated(EventSource source)
{
    if (!source.Name.Equals("System.Runtime"))
    {
        return;
    }

    EnableEvents(source, EventLevel.Verbose, EventKeywords.All, new Dictionary<string, string>()
    {
        ["EventCounterIntervalSec"] = "1"
    });
}

Sample code

Here is a sample EventListener class that prints all the counter names and values from the .NET runtime's EventSource, for publishing its internal counters (System.Runtime) every second.

using System;
using System.Collections.Generic;
using System.Diagnostics.Tracing;

public class SimpleEventListener : EventListener
{
    public SimpleEventListener()
    {
    }

    protected override void OnEventSourceCreated(EventSource source)
    {
        if (!source.Name.Equals("System.Runtime"))
        {
            return;
        }

        EnableEvents(source, EventLevel.Verbose, EventKeywords.All, new Dictionary<string, string>()
        {
            ["EventCounterIntervalSec"] = "1"
        });
    }

    protected override void OnEventWritten(EventWrittenEventArgs eventData)
    {
        if (!eventData.EventName.Equals("EventCounters"))
        {
            return;
        }

        for (int i = 0; i < eventData.Payload.Count; ++ i)
        {
            if (eventData.Payload[i] is IDictionary<string, object> eventPayload)
            {
                var (counterName, counterValue) = GetRelevantMetric(eventPayload);
                Console.WriteLine($"{counterName} : {counterValue}");
            }
        }
    }

    private static (string counterName, string counterValue) GetRelevantMetric(
        IDictionary<string, object> eventPayload)
    {
        var counterName = "";
        var counterValue = "";

        if (eventPayload.TryGetValue("DisplayName", out object displayValue))
        {
            counterName = displayValue.ToString();
        }
        if (eventPayload.TryGetValue("Mean", out object value) ||
            eventPayload.TryGetValue("Increment", out value))
        {
            counterValue = value.ToString();
        }

        return (counterName, counterValue);
    }
}

As shown above, you must make sure the "EventCounterIntervalSec" argument is set in the filterPayload argument when calling EnableEvents. Otherwise the counters will not be able to flush out values since it doesn't know at which interval it should be getting flushed out.